Institutional Investors and Stock ReturnsVolatility:

Empirical Evidence from a NaturalExperiment

Martin T. BohlWestfalische Wilhelms-University Munster, Germany

Janusz BrzeszczynskiHeriot-Watt University Edinburgh, United Kingdom

Bernd WilflingWestfalische Wilhelms-University Munster, Germany

Date of this version: September 18, 2007

Abstract: In this paper, we provide empirical evidence on the impact of institutionalinvestors on stock market returns dynamics. The Polish pension system reform in 1999and the associated increase in institutional ownership due to the investment activities

of pension funds are used as a unique institutional characteristic. Performing a Markov-Switching-GARCH analysis we find empirical evidence that the increase of institutionalownership has temporarily changed the volatility structure of aggregate stock returns.

The results are interpretable in favor of a stabilizing effect on index stock returns

induced by institutional investors.

JEL-classification codes: C32, G14, G23

Keywords: Institutional Traders, Polish Stock Market, Pension Fund Investors, StockMarket Volatility, Markov-Switching-GARCH Model

Earlier versions of this paper were presented at the Midwest Finance Association 54th AnnualMeeting, Milwaukee, the Eastern Finance Association Annual Meeting 2005, Norfolk, the 12th GlobalFinance Conference, Trinity College Dublin and the International Conference on Finance, Universityof Copenhagen. Comments provided by Carol Alexander, Jerzy Gajdka, Steven Isberg, Mark Schaffer,seminar participants and an anonymous referee are gratefully acknowledged. The first two authorsthank the Alexander von Humboldt Foundation for financial support.

Corresponding author: Janusz Brzeszczynski, Department of Accountancy and Finance, Heriot-Watt University, Edinburgh, EH14 4AS, United Kingdom. Phone: ++ 44 1314513294, Fax: ++ 441314513296, E-mail: J.Brzeszczynski@hw.ac.uk.

Institutional Investors and Stock ReturnsVolatility:

Empirical Evidence from a NaturalExperiment

Abstract: In this paper, we provide empirical evidence on the impact of institutional

investors on stock market returns dynamics. The Polish pension system reform in1999 and the associated increase in institutional ownership due to the investment ac-

tivities of pension funds are used as a unique institutional characteristic. Performing

a Markov-Switching-GARCH analysis we find empirical evidence that the increase ofinstitutional ownership has temporarily changed the volatility structure of aggregate

stock returns. The results are interpretable in favor of a stabilizing effect on indexstock returns induced by institutional investors.

JEL-classification codes: C32, G14, G23

Keywords: Institutional Traders, Polish Stock Market, Pension Fund Investors, Stock

Market Volatility, Markov-Switching-GARCH Model

11 Introduction

The increase in the number of institutional investors trading on stock markets world-wide since the end of the 1980s has caused a rise in financial economists interest in

institutions impact on stock prices. In particular, there is the suggestion that insti-tutional traders destabilize stock prices due to their specific investment behavior andthereby induce autocorrelation and increase volatility of stock returns. Among others,

herding and positive feedback trading are the two main arguments put forward for thedestabilizing impact on stock prices induced by institutional investors. Consequently,

empirical investigations have focused on the question of whether institutional tradersexhibit these types of investment behavior.1

However, evidence in favor of herding and positive feedback trading does not nec-essarily imply that institutional traders destabilize stock prices. If institutions herd

and all react to the same fundamental information in a timely manner, then insti-tutional investors speed up the adjustment of stock prices to new information andthereby make the stock market more efficient. Moreover, institutional investors maystabilize stock prices, if they jointly counter irrational behavior in individual investorssentiment. If institutional investors are better informed than individual investors, insti-

tutions will likely herd to undervalued stocks and away from overvalued stocks. Such

herding can move stock prices towards rather than away from fundamental values.Similarly, positive feedback trading is stabilizing, if institutional traders underreact tonews (Lakonishok, Shleifer and Vishny, 1992).

Consistent with the above arguments, Cohen, Gompers and Vuolteenaho (2002) find

a stabilizing impact of institutions on US stock prices. Institutions respond to positivecash-flow news by buying stocks from individual investors, thus exploiting the less thanone-for-one response of stock prices to cash-flow news. Moreover, in case of a price in-

crease in the absence of any cash-flow news institutions sell stocks to individuals. Thefindings by Cohen, Gompers and Vuolteenaho indicate that institutional investors push

stock prices to fundamental values and, hence, stabilize rather than destabilize stockprices. Barber and Odean (2005) find for the US that individual investors displayattention-based buying behavior on days of abnormally high trading volume, on daysof extremely negative and positive one-day returns and when stocks are in the news.In contrast, institutional investors do not exhibit attention-based buying. While the

1 Evidence on institutions trading behavior can be found in, for example, Lakonishok, Shleifer andVishny (1992), Grinblatt, Titman and Wermers (1995), Sias and Starks (1997), Nofsinger and Sias(1999), Wermers (1999), Badrinath and Wahal (2002) and Griffin, Harris and Topaloglu (2003), Siasand Whidbee (2006) and Yan and Zhang (2007).

2behavior of individual investors may contribute to stock returns autocorrelation andvolatility, institutions may induce a stabilizing effect on stock price dynamics. Support-

ing evidence also comes from the literature on the trading behavior and the impactof foreign, predominantly institutional, investors. Choe, Kho and Stulz (1999) and

Karolyi (2002) analyze data during crisis periods from Korea and Japan, respectively.Both investigations conclude that although foreign investors appear to follow positivefeedback trading strategies their trading behavior does not destabilize the markets.

We can conclude from the short discussion above that empirical findings on insti-tutional investors herding and positive feedback trading behavior are not necessarily

evidence in favor of a destabilizing effect on stock prices. Hence, these results provideonly indirect empirical evidence on the destabilizing effects of institutional investorstrading behavior on stock prices. To our best knowledge no empirical evidence is avail-

able about the direct effect of institutional traders destabilizing impact on stock prices.The existing literature on institutional trading behavior is predominantly forced to rely

on quarterly ownership data to compute changes in institutional holdings and in turndraws conclusions about the behavior of institutional investors.2 In contrast, under

the condition that the entrance date of a large number of institutional investors in thestock market is known, a Markov-switching-GARCH model may provide direct empiri-cal evidence of whether institutions change significantly the volatility structure of stock

index returns. In a time series framework we are able to investigate empirically theconsequences of a structural break in institutional ownership on stock returns volatilitybehavior.

The short history of the Polish stock market provides a unique institutional featurewhich allows us to contribute to the literature on the institutional investors impact on

stock prices. The special characteristic arises from the pension system reform in Poland.In 1999 privately managed pension funds were established and allowed to invest on thecapital market. We focus on the volatility behavior of stock returns prior to and afterthe first transfer of money to the pension funds on 19 May 1999. The appearance of

large institutional traders and the resulting increase in institutional ownership allows

us to investigate the impact on stock returns volatility in the environment of a naturalexperiment. Specifically, we use a modification of the Markov-switching-GARCHmodel

put forward by Gray (1996a) to study whether the models key coefficients change afterthe entrance of institutional traders in the Polish stock market. The main advantage of

2Sias, Starks and Titman (2006) provide a solution to the problem that high-frequency institutionalownership data are not available. The estimates of higher frequency covariances between changes ininstitutional ownership and stock returns rely on the use of higher frequency stock returns data andthe exploitation of covariance linearity.

3this econometric method is that it does not require an exogenously predetermined datefor the shift in stock returns volatility. Instead, Markov-switching-GARCH models

allow for endogenous specifications of volatility regime shifts and thus let the dataspeak for themselves.

The remainder of the paper is organized as follows. Section 2 contains a brief de-scription of the pension system reform and its consequences for the investors structureon the stock market in Poland. In section 3, the time series methodology, data and the

empirical results are outlined. Section 4 summarizes and concludes.

2 Pension System Reform and Investors Structureon the Stock Market in Poland

Re-established in 1991 the Polish stock market has grown rapidly during the last decade

in terms of the number of companies listed and the market capitalization. In compar-ison to the two other EU accession countries in the region, i.e. Czech Republic andHungary, the capitalization of the Polish stock market is significantly higher. It is

comparable to the ones of smaller mature European markets, like the Austrian stockmarket, and equalled about 60 billion $-US at the end of 2004 (Warsaw Stock Exchange,2005).The major change in the investors structure on the Polish stock market has its

origin in the pension system reform. In 1999, the public system was enriched bya private component, represented by open-end pension funds. Participation in this

component is mandatory for the employees below certain age. They are obliged totransfer 7.3 % of their gross salary to the government-run social insurance institute

called Zaklad Ubezpieczen Spolecznych (ZUS), which in turn transfers it to the pension

funds. The first transfer of money from the ZUS to the pension funds took place on19 May 1999. This date changed the investors structure of the Polish stock marketsignificantly. In 1999, about 20% domestic institutional investors and 45% domesticindividual investors traded at the Warsaw Stock Exchange. Over time the proportionof domestic institutional traders has increased, whereas the relative importance of

individual investors has decreased. In 2004, approximately one-third of the investorswere domestic individuals, and about one-third were national institutions. Constantlyabout one-third of the investors on the Polish stock market adhere to the group offoreign investors. The growing importance of pension fund investors is also reflectedby a gradual increase in annual ZUS transfers invested on the Warsaw Stock Exchangein relation to the average daily trading volume. This ratio increased from below 200%

4in 1999 to about 1000% in 2002 and has remained approximately constant since then.While before 19 May 1999 the majority of traders were small, private investors, after

that date pension funds became important players on the stock market. There were alsosome mutual funds active in the market but they had relatively small amounts of capital

under management. Moreover, the role of corporate investors was very marginal. It isthis feature in the history of the Polish stock market which constitutes the major changein the investors structure. This unique institutional characteristic allows us to compare

the period before 19 May 1999 characterized mainly by non-institutional trading withthe period after that date, where pension funds act as institutional investors on the

stock market. For reasons of argument it is important to stress that around this datethere were no other stock-market features which were of comparable importance as themarket entrance of the Polish pension funds.

The number of pension funds in 1999 2003 varied between 15 and 21. The changein their number occurred mainly due to the acquisitions of smaller funds by larger

ones. By the end of 2003, 17 pension funds operated in the Polish stock market withabout 12 billion $-US under management. In comparison, Polish insurance companiesand mutual funds had only 3 and 1 billion $-US of assets, respectively. In 2003, the

pension funds invested about 4 billion $-US in stocks listed on the Warsaw StockExchange. Their stock holdings predominantly consist of large capitalization stocks

that are listed in the blue-chip index WIG20 and usually belong to the Top 5 in theirindustries. Therefore, since May 1999 pension funds are important players on thePolish stock market, able to affect stock prices. In addition to their role as investors onthe stock market, Polish pension funds gained significant control in companies quoted

on the Warsaw Stock Exchange and executed their shareholder rights by appointing

members of the supervisory boards.Before May 1999, primarily individual private investors populated the Polish stock

market. The stock market was re-opened in 1991 after being closed for nearly half a

century. Thus, stock trading created a new investment opportunity for domestic private

investors and attracted many individuals who, in a very short period of time, openednearly 1 million brokerage accounts. While the level of the earliest available index

WIG remained far below 2000 points in the period from September 1991 to beginning

of May 1993, it jumped to 2027.7 on 6 May 1993 and reached the level of 20760.30 on 8March 1994, an increase of 924% within 10 months. Anecdotal evidence indicates thattrading decisions by Polish individual investors during this period often relied on non-professional information sources and gossips which, in turn, led to herding behavior.

An indicator of lack of fundamentally relevant information on companies listed at the

5Warsaw Stock Exchange is the limited access to information from professional dataproducers. The Reuters domestic news service for individual investors, Reuters Serwis

Polski, was introduced in Poland in the late 1990s and Reuters competitors followedwith their products even later in the early 2000s.

3 Econometric analysis

3.1 Data

Our data set consists of daily close prices of the Polish stock market index WIG20 andthe US index S&P500 covering the period between 1 November 1994 and 30 December

2003. The sample begins with the first complete month during which trading took place

5 days a week. Choosing 30 December 2003 as the end of the sample, we have the samenumber of months before and after the event in May 1999. Both indices were collectedfrom Datastream. The WIG20 index is selected because it contains the 20 largest

Polish stocks which are primarily held by the pension funds. Hence, using the WIG20

we approximate the pension funds portfolio composition. Figure 1 displays both indextime series where the stock return is defined as Rt = 100 ln(indext/indext1).

Figure 1 about here

As can be seen in Figure 1, the Polish stock market experienced a bull market sincethe mid 1990s. Unlike the US market, the Polish up market was interrupted by a

downturn in the second half of 1998 but recovered quite quickly until the beginning of2000. After the entrance of pension funds in May 1999 Polish stock prices increased

moderately until mid July and decreased gradually until November 1999. Then stock

prices increased drastically and reached their highest level in March 2000. Starting inApril 2000 and ending in October 2001 Polish stock prices declined for a relatively long

period. When looking at the graph for stock returns, Polish index returns show thewell-known volatility clustering.

3.2 A Markov-Switching-GARCH model

An appropriate econometric technique for analyzing stochastic volatility shifts is pro-vided by Markov-switching-GARCH models. Apart from some early methodological

contributions to Markov-switching models scattered in the literature, their modernformal foundation is due to Hamilton (1988, 1989). In our analysis we make use of a

6Markov-switching-GARCH model as developed in Gray (1996a), but modify his frame-work in two respects. First, we adapt Grays model for t-distributed index returns

within each regime and second, we incorporate a GARCH-dispersion specification as

proposed by Dueker (1997).3

The idea of an univariate Markov-switching model is that the data generating process

of the variable of interesthere of the daily stock returns of the WIG20 indexmaybe affected by a non-observable random variable St which represents the state the data

generating process is in at date t. In our analysis, the state variable St differentiatesbetween two volatility regimes and consequently takes on two distinct values. St = 1

indicates that the data generating process of the WIG20 index returns is in the high-volatility regime whereas for St = 2 the generating process is in the low volatilityregime.To set up our Markov-switching-GARCH model, recall first the probability density

function of a (displaced) t-distribution with degrees of freedom, mean and varianceh:

t,,h(x) =[( + 1)/2]

[/2] pi ( 2) h [1 +

(x )2h ( 2)

](+1)/2, (1)

where (z) 0 tz1 et dt, z > 0, denotes the complete gamma function. Next,we will specify stochastic processes for the mean and the variance in regime i (it and

hit, respectively) according to which the return at date t (denoted by Rt) is gener-

ated conditional upon the regime indicator St = i, i = 1, 2. Following Grays (1996a)

Markov-switching framework, the conditional distribution of the returns can be repre-sented as a mixture of two displaced t-distributions:

Rt|t1 t1,1t,h1t with probability p1tt2,2t,h2t with probability (1 p1t) , (2)

where t represents the usual time-t information set and p1t Pr {St = 1|t1} denotesthe so-called ex-ante probability of being in regime 1 at time t.

In our regime-dependent mean equations we explicitly take into account the possi-bility of first order autocorrelation in stock returns (by including Rt1) and the inter-dependence of the Polish stock market with the international stock market. For this

latter aspect we include the lagged S&P500 index returns RSPt1 as a control variable in3The use of t-distributed rather than normally distributed returns within each regime is motivated

by the fat-tail-property of stock index returns (Bollerslev, 1987). Alternatively, any other heavy-tailed parametric distribution like the Generalized Error Distribution (GED) suggested by Nelson(1991) could be specified to govern the tail thickness of the index returns. However, as will be arguedbelow, the t-distribution constitutes a good empirical model for our dataset.

7the mean equation:

it = a0i + a1i Rt1 + a2i RSPt1 for i = 1, 2. (3)

In contrast to the mean equation (3) the specification of an adequate GARCH-process for the regime-specific variance hit is more problematic. Technically, this com-plication is phrased as path dependence and stems from the GARCH lag structurewhich causes the regime-specific conditional variance to depend on the entire history

{St, St1, . . . , S0} of the regime-indicator St. We will circumvent this problem by ap-plying the same collapsing procedure as Gray (1996a). For this we have posited in

Eq. (2) that the data generating process that determines which regime observation t

comes from in fact depends on the probability pit as calculated from Eq. (9) below.

From Eq. (2) the variance of the stock return at date t can be expressed as:

ht = E[R2t |t1

] {E [Rt|t1]}2= p1t

(21t + h1t

)+ (1 p1t)

(22t + h2t

) [p1t 1t + (1 p1t) 2t]2 . (4)The quantity ht can be thought of as an aggregate of conditional variances from

both regimes and now provides the basis for the specification of the regime-specific

conditional variances hit+1, i = 1, 2 in the form of parsimonious GARCH(1,1) models.However, instead of using a conventional GARCH(1,1) structure, we follow the econo-metric motivation by Dueker (1997) and adopt a slightly modified GARCH equation.

For this, it is convenient to parameterize the degrees of freedom from the t-distribution

(1) by q = 1/, so that (1 2q) = ( 2)/, and to specify the alternative GARCHequation as:

hit = b0i + b1i (1 2qi) 2t1 + b2i ht1 (5)with ht1 as given according to Eq. (4), while t1 is obtained from:

t1 = Rt1 E [Rt1|t2]= Rt1 [p1t1 1t1 + (1 p1t1) 2t1] . (6)

To close the model, it remains to specify the transition probabilities of the regime

indicator St. For simplicity we consider a first order Markov process with constant

8transition probabilities, i.e. for pi1, pi2 [0, 1] we define:Pr {St = 1|St1 = 1} = pi1,Pr {St = 2|St1 = 1} = 1 pi1,Pr {St = 2|St1 = 2} = pi2,Pr {St = 1|St1 = 2} = 1 pi2.

(7)

Now, following Wilfling (2007), we obtain the log-likelihood function of our

Markov-switching-GARCH(1,1) model:

=Tt=1

log

p1t [(1 + 1)/2][1/2] pi 1 h1t [1 +

(Rt 1t)2h1t 1

](1+1)/2

+(1 p1t) [(2 + 1)/2][2/2] pi 2 h2t [1 +

(Rt 2t)2h2t 2

](2+1)/2 . (8)The log-likelihood function (8) contains the ex-ante probabilities p1i = Pr{St = 1|t1}.The whole series of ex-ante probabilities can be estimated recursively by

p1t = pi1 f1t1p1t1f1t1p1t1 + f2t1 (1 p1t1) + (1 pi2) f2t1 (1 p1t1)

f1t1p1t1 + f2t1 (1 p1t1) , (9)

where f1t and f2t denote the t1,1t,h1t- and t2,2t,h2t-density functions from Eq. (1),each evaluated at x = Rt.

Table I about here

3.3 Empirical results

Table I presents the maximum-likelihood estimates of the Markov-switching-GARCHmodel from the Eqs. (1) to (9) for the WIG20 index returns. The model was estimatedusing the full dataset covering 2391 trading days between 1 November 1994 and 30

December 2003 as described above. Furthermore, we investigated a shortened datasetconsisting of 392 trading days between 1 September 1998 and 1 March 2000 to take

into account the effect of major financial crises on Polish stock returns. The shortsample starts after the Asian and Russian crisis and ends before the world-wide collapsof stock markets in early 2000. Hence, we exclude the possibility that financial crises

before May 1999 are responsible for the relative lower volatility in stock returns after

the entrance of pension funds investors in the stock market. Maximization of the log-

9likelihood function was performed by the MAXIMIZE-routine within the softwarepackage RATS 6.01 using the BFGS-algorithm, heteroscedasticity-consistent estimates

of standard errors and suitably chosen starting values for all parameters involved.The estimates in Table I can be analyzed and interpreted economically. Before

analyzing the coefficients of the mean and GARCH equations (3) and (5), four aspectsof model specification and model diagnostics are worth mentioning.

(a) The first specification issue concerns the functional form of the GARCH equa-tion (5). Finance theory and empiricism suggest a positive relationship betweenthe perceived risk of an asset and its return on average. Within a single-regimetime-series framework this and other asymmetry considerations have led to vari-ous refined and mostly nonlinear GARCH specifications such as the GARCH-M

model (Engle et al., 1987), the EGARCHmodel (Nelson, 1991) and the TGARCHspecification (Glosten et al., 1993; Zakoian, 1994). Within a Markov-switching

framework with two (or even more) regimes, however, at least parts of these asym-metries and relationships may be captured by ordinary linear autoregressive and

GARCH specifications of the distinct regime-specific mean- and volatility equa-

tions. For example, the two regime-specific mean equations given by Eq. (3) inconjunction with the two regime-specific volatility equations from Eq. (5) arewell-suited to capture the empirically frequently-encountered finding that higher

perceived risk should pay a higher return on average. In such a situation thehigh-volatility (low-volatility) regime would be linked with that mean equation

which generates the higher (lower) mean returns. However, in some situations itmight appear appropriate to specify nonlinear GARCH equations (like EGARCHor TGARCH) for each individual regime. In principle, our Markov-switchingframework can be extended to accommodate these and even more general asym-metric GARCH specifications (e.g. those examined by Hentschel, 1995) in eachregime. Unfortunately, up to now the econometric properties of the resultingMarkov-switching GARCH-M, EGARCH or TGARCH models have not yet beenexplored. Since a rigorous mathematical analysis with respect to estimation,

hypothesis-testing and specification issues for these new model classes is beyondthe scope of this paper, we stick to safe econometric ground and use the linearGARCH specification (5) in our Markov-switching framework.

(b) Another specification issue concerns the statistical significance of the second

Markov regime as opposed to a single-regime GARCH specification. Unfortu-nately, a conventional likelihood ratio test (LRT) for testing the significance ofthe second regime turns out to be statistically improper, since under our model

10

setup there are seven parameters which remain unidentified under the null hy-pothesis of a single regime (i.e. when pi1 = 1 and pi2 = 0). However, bearing inmind the statistical dubiousness of the LRT, we follow a frequently encounteredapproach and report the values of the conventional LRT-statistics here.4 For this

purpose, we fitted single-regime models with mean and GARCH specifications

analogous to our equations (3) and (5) whose log-likelihood values are given inTable I (row Log-Likelihood: One-regime model). In conjunction with the log-likelihood values of our Markov-switching-GARCH model (row Log-Likelihood:

Two-regime model in Table I) we computed the LRT statistics for both datasetsas twice the difference between the log-likelihood values of the respective spec-

ifications (row LRT in Table I). If the LRT were statistically valid, we wouldhave had to compare the LRT statistics against the critical values derived fromthe quantiles of a 2-distribution with nine degrees of freedom (since the two-regime specification has 9 additional parameters as opposed to the single-regimemodel). Now, the critical value of a 2(9)-distribution at the 1% level is 21.6660.Obviously, the LRT statistics for both datasets clearly exceed this critical value

providing at least some (statistically informal) confidence in the existence of asecond regime.5

(c) All q-parameters except for q1 (i.e. for Regime 1) of the shortened dataset arelarger than zero at any conventional significance level. It is well-known that the

t-distribution (1) converges to the normal distribution for q = 1/ 0, but hasfatter tails than the corresponding normal distribution for any finite . This

implies significant deviations from the normal distribution for the Regimes 1and 2 of the full dataset and for Regime 2 of the shortened dataset. Moreover,the estimates of the degree-of-freedom parameters 1 = 1/q1 and 2 = 1/q2are all larger than 4.0, explicitly ranging between 6.5359 and 344.8276. Thisresult has two important implications for all regime-specific (time-varying) t-

4This simplifying approach has been adopted among others by Hamilton and Susmel (1994) andGray (1996a). Ang and Bekaert (2002) provide a statistically more stringent justification for thisapproach.

5The modelling of only two Markov regimes, namely a low- and a high-volatility regime, mightappear unrealistic at first glance. The consideration of at least one additional intermediate volatil-ity regime seems natural. Unfortunately, the estimation of Markov-switching models with three oreven more regimes becomes numerically unfeasible due to an exploding number of parameters arisingfrom each additional Markov regime included in the econometric specification (see Wilfling, 2007, forstatistical details). However, since we are merely interested in testing for an overall-effect on stock-return volatility (i.e. either for an overall-volatility increase or for an overall-volatility decrease) causedby the institutional investors entrance into the Polish stock market, a two-regime Markov-switchingmodelapart from being numerically estimablealso appears to represent a factually well-groundedspecification.

11

distributions estimated on the basis of our dataset. First, all t-distributions havefinite variances (Hamilton, 1994). Second, all t-distributions have finite kurtosis(Mood et al., 1974). Obviously, the t-distribution constitutes a highly adequateempirical specification to capture the fat-tail property of stock-index returns for

our dataset.

(d) The lower part of Table I contains a diagnostic check of the model fit by pro-viding Ljung-Box statistics for serial correlation of the squared (standardized)residuals out to the lags 1, 2, 3, 5, 10. Obviously, the null hypothesis of noautocorrelation cannot be rejected out to all lags at any conventional signifi-

cance level providing further econometric evidence in favour of our two-regimeMarkov-switching-GARCH specification.

The majority of the estimated coefficients of the mean and GARCH equations (3)and (5) are statistically significant at the 1% level. The autoregressive coefficientsa11 for regime 1 are statistically significant and negative for both datasets while thecoefficients a12 for regime 2 are positive for both datasets. It is informative to notethat the negative autoregressive coefficients a11 are contradictory to a result often

reported in the literature finding a positive autoregressive structure of order one in

stock index returns due to non-synchronous trading (Lo and MacKinlay, 1990), time-varying expected returns (Conrad and Kaul, 1988) and transaction costs (Mech, 1993).

The coefficients of the control variable RSPt1 are statistically significant (at least at the5% level) and positive in both regimes revealing the strong interdependence betweenUS and Polish stock returns dynamics.

When looking at the estimated parameters describing the conditional volatility pro-cess we find the well-established result of volatility persistence for both datasets and

in both regimes (except for regime 2 in the shortened dataset). However, none of thefour coefficient sums b1i (1 2qi) + b2i, i = 1, 2, for both datasets exceeds 1 providingsome evidence for stationary conditional volatility processes in all regimes.

The constant transition probabilities pi1 and pi2 are close to one in both estimations.Since both quantities represent the probability of the data generating process remaining

in the same volatility regime during the transition from date t 1 to t, both volatilityregimes reveal a high degree of persistence.

Next, we address two conditional probabilities which are of inferential relevance fordetecting how often and at which dates the Polish stock market switched between the

high and the low volatility regimes. First, the ex-ante probabilities p1t Pr{St =1|t1}, t = 2, . . . , T , which can be estimated recursively via Eq. (9), and second, the

12

so-called smoothed probabilities Pr{St = 1|T}, t = 1, . . . , T , which can be computedafter model estimation by the use of filter techniques.6

The ex-ante probabilities are useful in forecasting one-step-ahead regimes based on aninformation set which evolves over time. In our context, the ex-ante probabilities reflectcurrent market perceptions of the one-step-ahead volatility regime, thus representingan adequate measure of stock market volatility sentiments. In contrast to this, thesmoothed probabilities are based on the full sample-information set T and thus providea basis for inferring ex post if and when volatility regime switches have occurred in the

sample.

Figures 2 and 3 about here

Figures 2 and 3 display both regime-1 probabilities (upper panels) along with the con-

ditional variance processes (lower panels) estimated for the Markov-switching-GARCHmodel on the basis of the full and the shortened datasets, respectively. The ex-ante

probabilities are represented by the thin lines while the bold lines depict the smoothedregime-1 probabilities. Since the ex-ante probabilities are determined by an evolving

(and thus smaller) information set, they exhibit a more erratic dynamic behaviour than

the smoothed regime-1 probabilities. As a visual support, all panels contain a marker

for the 19 May 1999, the crucial date of the Polish pension system reform.

As can be seen in Figure 2, after a period of low conditional volatility we observea jump to a high volatility regime around February 1997 when all blue chip stockscontained in the WIG20 became continuously traded. High conditional variances at

the end of 1997, mid 1998 and at the beginning of 1999 are associated with the Asian(October/November 1997), the Russian (August/September 1998) and the Brazilian

(January 1999) crises, respectively. More importantly, Figures 2 and 3 demonstrate theeffect of the change in the investors structure in the Polish stock market on 19 May1999. The conditional volatility process exhibits a structural break around this date.While the conditional variances are higher before May 1999, they are significantly lowerafterwards in both Figures. Further econometric evidence is provided by the ex-anteand the smoothed probabilities in the upper panels which show a clear-cut transition

from a high to a low volatility regime around the date of the entrance of pension

fund investors. In 2000 the low volatility regime switches again to the high volatilityregime. We can conclude that the entrance of institutional investors on the Polishstock market reduced at least temporarily the volatility of stock returns. While the

6The smoothed probabilities for the WIG20 index returns were computed on the basis of a filteralgorithm provided by Gray (1996b).

13

high volatility regime in 2000 can be explained by the bear market, the evidence around19 May 1999 convincingly demonstrates the stabilizing effect of institutional investors

on Polish stock price dynamics.

4 Summary and Conclusions

One of the most prominent changes in financial markets during the recent decadesis the surge of institutional investors. Concerning their specific investment behavior

numerous studies indicate that institutional investors engage in herding and tend toexhibit positive feedback trading strategies and thus contribute to stock returns auto-correlation and volatility. In this paper, we challenge this view and provide empirical

evidence on the influence of institutional investors on stock returns dynamics. ThePolish pension reform in 1999 is used as an institutional peculiarity to implement aMarkow-switching-GARCH model. Before and after 19 May 1999 there were no other

features of the Polish stock market which were of comparable relevance as the marketentrance of pension funds. Therefore, it is this institutional feature of the Polish emerg-

ing stock market together with the econometric technique that allows us to answer the

following questions: Did the increase of institutional ownership after the appearanceof Polish pension funds on 19 May 1999 result in a change in the volatility structureof stock index returns? Did Polish pension fund investors destabilize or stabilize stockprices?

We provide empirical evidence in favor of a change in the conditional volatility

process due to the increased importance of institutional investors on the Polish stockmarket. In contrast to the often mentioned suggestion that institutional investorsincrease stock returns volatility, our findings support the hypothesis that the pension

fund investors in Poland reduced stock market volatility. Hence, our empirical evidenceis in favor of a stabilizing rather than a destabilizing effect induced by pension fundsinvestors in Poland.In a broader perspective our findings are supportive of the view that institutional

investors can be characterized as informed investors who speed up the adjustment ofstock prices to new information thereby making the stock market more efficient. In-

stitutions can create an informational advantage by exploiting economies of scale ininformation acquisition and processing. The marginal costs of gathering and processingare lower than for individual traders. In this sense our findings are consistent with the

evidence in Dennis and Weston (2001) for the US. If individual investors contribute to

stock returns volatility, a significant decrease in trades by individuals relative to insti-

14

tutions might provide an explanation for the stabilizing effect. Moreover, institutionalinvestors may stabilize stock prices and counter irrational behavior in individual in-

vestors sentiment. Gabaix et al. (2006) provide a theoretical model in which trades bylarge institutional investors in relatively illiquid markets generate excess stock market

volatility. Our empirical findings do not support this theoretical prediction.

15

References

Ang, A., Bekaert, G., 2002. Regime Switches in Interest Rates, Journal of Business

and Economic Statistics 20, 163 - 182.Badrinath, S. G., Wahal, S., 2002. Momentum Trading by Institutions, Journal of

Finance 57, 2449 - 2478.Barber, B. M., Odean, T., 2005. All that Glitters: The Effect of Attention and News

on the Buying Behavior of Individual and Institutional Investors, Working Paper.(Available under: http://ssrn.com.abstract=460660.)

Bollerslev, T., 1987. A Conditionally Heteroskedastic Time Series Model for Specula-tive Prices and Rates of Return, The Review of Economics and Statistics 69, 542

- 547.Choe, H., Kho, B.-Ch., Stulz, R., 1999. Do Foreign Investors Destabilize Stock Mar-

kets? The Korean Experience in 1997, Journal of Financial Economics 54, 227 -264.

Cohen, R. B., Gompers, P. A., Vuolteenaho, T., 2002. Who Underreacts to Cash-FlowNews? Evidence from Trading between Individuals and Institutions, Journal of

Financial Economics 66, 409 - 462.Conrad, J., Kaul, G. 1988. Time-Variation in Expected Returns, Journal of Business

61, 409 - 25.Dennis, P. J., Weston, J. P., 2001. Whos Informed? An Analysis of Stock Ownership

and Informed Trading, Working Paper.

(Available under: http://ssrn.com.abstract=267350.)Dueker, M.J., 1997. Markov Switching in GARCH Processes and Mean-Reverting

Stock Market Volatility, Journal of Business and Economic Statistics 15, 26 - 34.Engle, R.F., Lilien, D.M., Robins, R.P., 1987. Estimating Time Varying Risk Premia

in the Term Structure: The ARCH-M Model, Econometrica 55, 391-407.Gabaix, X., Gopikrishnan, P., Plerou, V., Stanley, H.E., 2006. Institutional Investors

and Stock Market Volatility, Quarterly Journal of Economics 121, 461 - 504.Glosten, L.R., Jagannathan, R., Runkle, D.E., 1993. On the Relation Between the

Expected Value and the Volatility of the Nominal Excess Return on Stocks,

Journal of Finance 48, 1779-1801.Gray, S.F., 1996a. Modeling the Conditional Distribution of Interest Rates as a

Regime-Switching Process, Journal of Financial Economics 42, 27 - 62.Gray, S.F., 1996b. An Analysis of Conditional Regime-Switching Models. Working

paper, Fuqua School of Business, Duke University, Durham NC.Griffin, J. M., Harris, J. H., Topaloglu, S., 2003. The Dynamics of Institutional and

Individual Trading, Journal of Finance 58, 2285 - 2320.

16

Grinblatt, M., Titman, S., Wermers, R., 1995. Momentum Investment Strategies, Port-folio Performance, and Herding: A Study on Mutual Fund Behavior, American

Economic Review 85, 1088 - 1105.Hamilton, J.D., 1988. Rational-Expectations Econometric Analysis of Changes in

Regime: An Investigation of the Term Structure of Interest Rates, Journal of

Economic Dynamics and Control 12, 385 - 423.Hamilton, J.D., 1989. A New Approach to the Economic Analysis of Nonstationary

Time Series and the Business Cycle, Econometrica 57, 357-384.Hamilton, J.D., 1994. Time Series Analysis. Princeton University Press, Princeton.Hamilton, J.D., Susmel, R., 1994. Autoregressive Conditional Heteroskedasticity and

Changes in Regime, Journal of Econometrics 64, 307-333.Hentschel, L.E., 1995. All in the Family: Nesting Symmetric and Asymmetric GARCH

Models, Journal of Financial Economics 39, 71-104.Karolyi, A., 2002. Did the Asian Financial Crisis Scare Foreign Investors out of Japan?,

Pacific-Basin Finance Journal 10, 411 - 442.Lakonishok, J., Shleifer, A., Vishny, R. W., 1992. The Impact of Institutional Trading

on Stock Prices, Journal of Financial Economics 32, 23 - 43.Lo, A., MacKinlay, A. C., 1990. An Econometric Analysis of Non-Synchronous Trad-

ing, Journal of Econometrics 45, 181 - 212.Mech, T., 1993. Portfolio Return Autocorrelation, Journal of Financial Economics 34,

307 - 44.Mood, A.M., Graybill, F.A., Boes, D.C., 1974. Introduction to the Theory of Statistics.

Third Edition, McGraw-Hill, Tokyo.Nelson, D.B., 1991. Conditional Heteroskedasticity in Asset Returns: A New Ap-

proach, Econometrica 59, 347-370.Nofsinger, J. R., Sias, R. W., 1999. Herding and Feedback Trading by Institutional

and Individual Investors, Journal of Finance 54, 2263 - 2295.Sias, R. W., Starks, L. T., 1997. Return Autocorrelation and Institutional Investors,

Journal of Financial Economics 46, 103 - 121.Sias, R.W., Starks, L.T., Titmann, S., 2006. Changes in Institutional Ownership and

Stock Returns: Assessment and Methodology, Journal of Business 79, 2869 -

2910.Sias, R., Whidbee, D.A, 2006. Are Institutional or Individual Investors More Likely to

Drive Prices from Fundamentals, Working Paper, Washington State University.Wermers, R., 1999. Mutual Fund Herding and the Impact on Stock Prices, Journal of

Finance 54, 581 - 622.

17

Warsaw Stock Exchange, 2005. Fact Book, Warsaw, Poland.(Available under: www.wse.com.pl.)

Wilfling, B., 2007. Volatility Regime-Switching in European Exchange Rates prior toMonetary Unification, Journal of International Money and Finance, forthcoming.

Yan, X., Zhang, Z., 2007. Institutional Investors and Equity Returns: Are Short-TermInstitutions Better Informed?, Review of Financial Studies, forthcoming.

Zakoian, J.M., 1994. Threshold Heteroskedastic Models, Journal of Economic Dynam-ics and Control 18, 931-944.

18

Table IEstimates and related statistics for Markov-switching-GARCH models

Full dataset Shortened datasetEstimate Std. error Estimate Std. error

Regime 1:a01 0.0643 0.0623 0.2120 0.1830a11 0.0952** 0.0316 0.2234** 0.0523a21 0.5300** 0.0400 0.9648** 0.1438b01 0.1439 0.0819 3.0571** 0.1174b11 0.1317** 0.0272 0.1101** 0.0023b21 0.8282** 0.0410 0.8502** 0.0550q1 = 1/1 0.1333** 0.0218 0.0029 0.0550[b11 (1 2q1) + b21] [0.9248] [0.9597]Regime 2:a02 0.0395 0.0356 0.0572 0.1011a12 0.0909** 0.0291 0.0143 0.0675a22 0.1689** 0.0304 0.1672* 0.0729b02 0.0764* 0.0355 1.3128** 0.2562b12 0.0882** 0.0197 0.1188* 0.0481b22 0.8692** 0.0348 0.0614 0.1259q2 = 1/2 0.1530** 0.0282 0.1278** 0.0403[b12 (1 2q2) + b22] [0.9304] [0.1498]Transition probabilities:pi1 0.9982** 0.0023 0.9826** 0.0045pi2 0.9991** 0.0006 0.9963** 0.0018Log-Likelihood:Two-regime model 4721.6049 781.9689One-regime model 4739.8082 794.8619LRT 36.4066 25.7860Residual analysis Test statistic p-value Test statistic p-valueLB21 0.6602 0.4165 0.0801 0.7771LB22 0.9398 0.6251 0.2341 0.8895LB23 1.5727 0.6656 0.3556 0.9492LB25 3.7951 0.5793 0.4368 0.9943LB210 12.4444 0.2564 5.8881 0.8246Note: Estimates for parameters from the Eqs. (1) to (9). LB2i denotes the Ljung-Box-Q-statisticfor serial correlation of the squared standardized residuals out to lag i. ** and * denote statisticalsignificance at the 1 % and 5 % levels, respectively.

19

Figure 1: Stock market indexes and returns

20

Figure 2: Regime-1 probabilities and conditional variances (full dataset)

21

Figure 3: Regime-1 probabilities and conditional variances (shortened dataset)